Signal #59: Fable Comes In From the Cold

Hello friendly subscribers

Three weeks ago the US government made Anthropic's strongest models illegal to export. This week it reversed itself, and Fable 5 starts returning to Europe on July 1. The timing is almost comic: the same stretch that freed Fable, Anthropic also shipped Sonnet 5, the "everyman" model that landed with a shrug. So this issue is built around the model you can suddenly reach again and how to get real work out of it.

Below: why the Fable ban ended, why Sonnet 5 impressed no one, a proper playbook for running Fable 5 in loops and workflows, and a pattern in CEE where fintechs are quietly building the rails AI agents will pay through.

On June 30 the US lifted its license requirement on Anthropic's Mythos and Fable models, and Anthropic said it would begin restoring access on July 1. The rule had existed only since June 12, when the government added both to its export-restricted list. Complying at scale proved impossible, so Anthropic pulled public access to what many rate as the most capable models yet shipped.

Read the reversal for what it says about who controls the frontier. Cybersecurity researchers doubted the ban was ever about security, since Anthropic had already pledged the safety steps voluntarily; it read more like a pressure tactic after the company's executives criticised the administration. What actually moved Washington was competition: Asian labs started shipping Mythos-level models like Fugu and Tulonfeng, and a ban that handicapped US labs stopped being tenable.

For anyone building outside the US, the most capable public model is legal to use again, so move your critical work back and test it this week. Do it with one eye open: a gate that closed on twelve days' notice can close again. Keep a second model wired into anything you can't afford to lose access to.

Most SaaS finance teams know their billing process isn't perfect. Few know what it's actually costing them.

Answer 5 quick questions — contracts signed per month, ACV, days to first invoice, error rate, DSO — and the Tabs Billing Lag Calculator gives you a dollar figure benchmarked against top SaaS companies.

It takes two minutes. The number might surprise you.

Calculate your billing lag and see where you stand.

Calculate your billing lag

Anthropic pitched Sonnet 5 as its most agentic Sonnet yet, near Opus 4.8 quality at a fraction of the price. Every's testers took it apart. Turned up to full effort it only matches Opus 4.8 running at low-to-medium, while costing more per task. Any speed edge vanishes into the time you spend fixing its missteps, and Kieran Klaassen watched it spend three hours looping on a broken build without solving anything.

The Goldilocks pitch only holds if the model is not too expensive, slow, or dull next to the alternatives, and Sonnet 5 is all three. Every reviewer found a better option for their job: Gemini at $2 per million tokens against Sonnet's $3, GPT-5.5 for writing, Opus and Fable for the hard work. The signal for operators is that "which model" is now a per-task decision, not a default you set once and forget.

Don't move a daily workflow onto Sonnet 5 because the number went up; benchmark it against your current default on one real task first. Note the tell buried in the reviews: testers reached back for Fable the moment it came online. Which is the model worth learning properly.

Fable 5 is a Mythos-class model built for long, autonomous runs, not for chat. Its edge is per-step accuracy, and in a workflow small errors compound, so higher accuracy on each tool call turns into a much better chance the whole job finishes clean. That means the way you get value out of Fable is different from how you prompt a chat model. You design the run. Here is the shape of it.

Stop prompting, start looping. For real work, don't write one clever prompt and hope. Define the outcome, the constraints, and the success criteria, then let Fable execute in a loop using its tools, checking its own output against those criteria and correcting until it clears the bar. The one move that matters most here is verification: Anthropic's engineers report that a separate verifier sub-agent beats self-critique by a wide margin, roughly six times more improvement than earlier models on some engineering tasks. Give the loop an honest rubric so it gets real feedback from the environment, not a model grading its own homework.

Give it memory across sessions. For anything that spans more than one sitting, have Fable write a STATE update before it ends a session and read that STATE at the start of the next one. That one habit is what lets it resume a long project without you re-briefing it every morning, and it is where continual, multi-day work actually becomes practical.

Let the workflow build itself. Fable's dynamic workflows decide their structure at runtime, not when you design them. Point it at a big job and it can spawn many parallel sub-agents, each taking a discrete chunk, then fold the results back into one coherent output. Research sweeps, code migrations, and bulk document analysis are the obvious fits: work that is wide rather than deep, where dividing the load is the whole trick.

Ask for structure, always. Reliability comes from explicit output shapes. Ask for JSON when the result feeds a program, numbered steps with stated dependencies when you want a plan you can act on, and a fixed template when a summary will be piped into the next prompt. Loose prose breaks downstream automation; structured output is what keeps a long chain from drifting.

Route, don't marry. Fable is expensive for easy calls, so run a hybrid setup: a cheaper model for general tasks, Fable reserved for the genuinely hard 20%. Wiring both in also means the next policy shock, or the next ban, doesn't strand your pipeline on a single provider.

The move this week: take one multi-step task you still do by hand, write down what "done well" means for it, and let Fable run it in a loop with a verifier checking the result. If you want a running start on prompts, Every published a Fable 5 prompt library worth raiding.

A quiet pattern runs through this spring's CEE funding data. Poland's Paymove raised €2.12M to build a payments layer aimed squarely at autonomous AI agents. Bulgaria's Paypercut pulled €5M (€7M in total) to unify cross-border card, local, and buy-now-pay-later rails across eight CEE markets, with stablecoin corridors and its own AI agents on the roadmap. Warsaw's sunbay is building agents that chase overdue invoices on their own.

While the US argues about whether agents should be allowed to spend money, CEE fintechs are building the plumbing for when they do. The region's fragmentation is the reason it fits: a mesh of currencies and thin corridors like EUR-to-PLN and EUR-to-RON is exactly the friction an agent-payments layer exists to erase. This is a bet placed before the category has a name in the headlines, which is usually where the good regional entry points hide.

If you build in payments or commerce, treat "infrastructure for agent transactions" as a category forming in your backyard rather than a Silicon Valley abstraction. If you are raising, note that the framing is already pulling regional cheques. The teams wiring this now will own the corridors when agents start paying through them.

Four tools to install or test this week. Sales and marketing were quiet, so productivity carries two.

Productivity: Notion Agents run tasks on their own. Notion shipped autonomous Agents that work across your whole workspace for up to 20 minutes at a stretch, completing multi-step jobs instead of just drafting text. Hand one a recurring task you usually click through by hand, like turning a meeting note into a project with sub-tasks, and judge whether it finishes clean.

Productivity: ClickUp put an agent in every channel. ClickUp launched two AI agents with a full redesign: one lives in your channels answering questions from internal docs plus Drive and Gmail, the other, Brain, schedules meetings, builds tasks, and drafts reports. If you already run ClickUp, switch Brain on for one project before adding another tool.

Design: Canva AI 2.0 connects to your stack. Canva's new connectors wire its AI into Slack, Gmail, Drive, HubSpot, and Zoom, so it can turn a Zoom transcript into a recap or a customer email into an on-brand pitch deck without leaving Canva. Point it at one repetitive branded asset you rebuild every week.

Dev: Fable 5's prompt library is worth raiding. With Fable back online, Every published a Fable 5 prompt library of tested patterns for loops, verification, and structured output. Steal one loop pattern and adapt it to a task in your own stack rather than starting the prompt from scratch.

Next edition soon,
Çelik